1 /* 2 * Copyright (c) Meta Platforms, Inc. and affiliates. 3 * All rights reserved. 4 * 5 * This source code is licensed under both the BSD-style license (found in the 6 * LICENSE file in the root directory of this source tree) and the GPLv2 (found 7 * in the COPYING file in the root directory of this source tree). 8 * You may select, at your option, one of the above-listed licenses. 9 */ 10 11 #include "zstd_compress_internal.h" /* ZSTD_hashPtr, ZSTD_count, ZSTD_storeSeq */ 12 #include "zstd_fast.h" 13 14 static 15 ZSTD_ALLOW_POINTER_OVERFLOW_ATTR 16 void ZSTD_fillHashTableForCDict(ZSTD_MatchState_t* ms, 17 const void* const end, 18 ZSTD_dictTableLoadMethod_e dtlm) 19 { 20 const ZSTD_compressionParameters* const cParams = &ms->cParams; 21 U32* const hashTable = ms->hashTable; 22 U32 const hBits = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS; 23 U32 const mls = cParams->minMatch; 24 const BYTE* const base = ms->window.base; 25 const BYTE* ip = base + ms->nextToUpdate; 26 const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE; 27 const U32 fastHashFillStep = 3; 28 29 /* Currently, we always use ZSTD_dtlm_full for filling CDict tables. 30 * Feel free to remove this assert if there's a good reason! */ 31 assert(dtlm == ZSTD_dtlm_full); 32 33 /* Always insert every fastHashFillStep position into the hash table. 34 * Insert the other positions if their hash entry is empty. 35 */ 36 for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) { 37 U32 const curr = (U32)(ip - base); 38 { size_t const hashAndTag = ZSTD_hashPtr(ip, hBits, mls); 39 ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr); } 40 41 if (dtlm == ZSTD_dtlm_fast) continue; 42 /* Only load extra positions for ZSTD_dtlm_full */ 43 { U32 p; 44 for (p = 1; p < fastHashFillStep; ++p) { 45 size_t const hashAndTag = ZSTD_hashPtr(ip + p, hBits, mls); 46 if (hashTable[hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) { /* not yet filled */ 47 ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr + p); 48 } } } } 49 } 50 51 static 52 ZSTD_ALLOW_POINTER_OVERFLOW_ATTR 53 void ZSTD_fillHashTableForCCtx(ZSTD_MatchState_t* ms, 54 const void* const end, 55 ZSTD_dictTableLoadMethod_e dtlm) 56 { 57 const ZSTD_compressionParameters* const cParams = &ms->cParams; 58 U32* const hashTable = ms->hashTable; 59 U32 const hBits = cParams->hashLog; 60 U32 const mls = cParams->minMatch; 61 const BYTE* const base = ms->window.base; 62 const BYTE* ip = base + ms->nextToUpdate; 63 const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE; 64 const U32 fastHashFillStep = 3; 65 66 /* Currently, we always use ZSTD_dtlm_fast for filling CCtx tables. 67 * Feel free to remove this assert if there's a good reason! */ 68 assert(dtlm == ZSTD_dtlm_fast); 69 70 /* Always insert every fastHashFillStep position into the hash table. 71 * Insert the other positions if their hash entry is empty. 72 */ 73 for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) { 74 U32 const curr = (U32)(ip - base); 75 size_t const hash0 = ZSTD_hashPtr(ip, hBits, mls); 76 hashTable[hash0] = curr; 77 if (dtlm == ZSTD_dtlm_fast) continue; 78 /* Only load extra positions for ZSTD_dtlm_full */ 79 { U32 p; 80 for (p = 1; p < fastHashFillStep; ++p) { 81 size_t const hash = ZSTD_hashPtr(ip + p, hBits, mls); 82 if (hashTable[hash] == 0) { /* not yet filled */ 83 hashTable[hash] = curr + p; 84 } } } } 85 } 86 87 void ZSTD_fillHashTable(ZSTD_MatchState_t* ms, 88 const void* const end, 89 ZSTD_dictTableLoadMethod_e dtlm, 90 ZSTD_tableFillPurpose_e tfp) 91 { 92 if (tfp == ZSTD_tfp_forCDict) { 93 ZSTD_fillHashTableForCDict(ms, end, dtlm); 94 } else { 95 ZSTD_fillHashTableForCCtx(ms, end, dtlm); 96 } 97 } 98 99 100 typedef int (*ZSTD_match4Found) (const BYTE* currentPtr, const BYTE* matchAddress, U32 matchIdx, U32 idxLowLimit); 101 102 static int 103 ZSTD_match4Found_cmov(const BYTE* currentPtr, const BYTE* matchAddress, U32 matchIdx, U32 idxLowLimit) 104 { 105 /* Array of ~random data, should have low probability of matching data. 106 * Load from here if the index is invalid. 107 * Used to avoid unpredictable branches. */ 108 static const BYTE dummy[] = {0x12,0x34,0x56,0x78}; 109 110 /* currentIdx >= lowLimit is a (somewhat) unpredictable branch. 111 * However expression below compiles into conditional move. 112 */ 113 const BYTE* mvalAddr = ZSTD_selectAddr(matchIdx, idxLowLimit, matchAddress, dummy); 114 /* Note: this used to be written as : return test1 && test2; 115 * Unfortunately, once inlined, these tests become branches, 116 * in which case it becomes critical that they are executed in the right order (test1 then test2). 117 * So we have to write these tests in a specific manner to ensure their ordering. 118 */ 119 if (MEM_read32(currentPtr) != MEM_read32(mvalAddr)) return 0; 120 /* force ordering of these tests, which matters once the function is inlined, as they become branches */ 121 #if defined(__GNUC__) 122 __asm__(""); 123 #endif 124 return matchIdx >= idxLowLimit; 125 } 126 127 static int 128 ZSTD_match4Found_branch(const BYTE* currentPtr, const BYTE* matchAddress, U32 matchIdx, U32 idxLowLimit) 129 { 130 /* using a branch instead of a cmov, 131 * because it's faster in scenarios where matchIdx >= idxLowLimit is generally true, 132 * aka almost all candidates are within range */ 133 U32 mval; 134 if (matchIdx >= idxLowLimit) { 135 mval = MEM_read32(matchAddress); 136 } else { 137 mval = MEM_read32(currentPtr) ^ 1; /* guaranteed to not match. */ 138 } 139 140 return (MEM_read32(currentPtr) == mval); 141 } 142 143 144 /** 145 * If you squint hard enough (and ignore repcodes), the search operation at any 146 * given position is broken into 4 stages: 147 * 148 * 1. Hash (map position to hash value via input read) 149 * 2. Lookup (map hash val to index via hashtable read) 150 * 3. Load (map index to value at that position via input read) 151 * 4. Compare 152 * 153 * Each of these steps involves a memory read at an address which is computed 154 * from the previous step. This means these steps must be sequenced and their 155 * latencies are cumulative. 156 * 157 * Rather than do 1->2->3->4 sequentially for a single position before moving 158 * onto the next, this implementation interleaves these operations across the 159 * next few positions: 160 * 161 * R = Repcode Read & Compare 162 * H = Hash 163 * T = Table Lookup 164 * M = Match Read & Compare 165 * 166 * Pos | Time --> 167 * ----+------------------- 168 * N | ... M 169 * N+1 | ... TM 170 * N+2 | R H T M 171 * N+3 | H TM 172 * N+4 | R H T M 173 * N+5 | H ... 174 * N+6 | R ... 175 * 176 * This is very much analogous to the pipelining of execution in a CPU. And just 177 * like a CPU, we have to dump the pipeline when we find a match (i.e., take a 178 * branch). 179 * 180 * When this happens, we throw away our current state, and do the following prep 181 * to re-enter the loop: 182 * 183 * Pos | Time --> 184 * ----+------------------- 185 * N | H T 186 * N+1 | H 187 * 188 * This is also the work we do at the beginning to enter the loop initially. 189 */ 190 FORCE_INLINE_TEMPLATE 191 ZSTD_ALLOW_POINTER_OVERFLOW_ATTR 192 size_t ZSTD_compressBlock_fast_noDict_generic( 193 ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 194 void const* src, size_t srcSize, 195 U32 const mls, int useCmov) 196 { 197 const ZSTD_compressionParameters* const cParams = &ms->cParams; 198 U32* const hashTable = ms->hashTable; 199 U32 const hlog = cParams->hashLog; 200 size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1; /* min 2 */ 201 const BYTE* const base = ms->window.base; 202 const BYTE* const istart = (const BYTE*)src; 203 const U32 endIndex = (U32)((size_t)(istart - base) + srcSize); 204 const U32 prefixStartIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog); 205 const BYTE* const prefixStart = base + prefixStartIndex; 206 const BYTE* const iend = istart + srcSize; 207 const BYTE* const ilimit = iend - HASH_READ_SIZE; 208 209 const BYTE* anchor = istart; 210 const BYTE* ip0 = istart; 211 const BYTE* ip1; 212 const BYTE* ip2; 213 const BYTE* ip3; 214 U32 current0; 215 216 U32 rep_offset1 = rep[0]; 217 U32 rep_offset2 = rep[1]; 218 U32 offsetSaved1 = 0, offsetSaved2 = 0; 219 220 size_t hash0; /* hash for ip0 */ 221 size_t hash1; /* hash for ip1 */ 222 U32 matchIdx; /* match idx for ip0 */ 223 224 U32 offcode; 225 const BYTE* match0; 226 size_t mLength; 227 228 /* ip0 and ip1 are always adjacent. The targetLength skipping and 229 * uncompressibility acceleration is applied to every other position, 230 * matching the behavior of #1562. step therefore represents the gap 231 * between pairs of positions, from ip0 to ip2 or ip1 to ip3. */ 232 size_t step; 233 const BYTE* nextStep; 234 const size_t kStepIncr = (1 << (kSearchStrength - 1)); 235 const ZSTD_match4Found matchFound = useCmov ? ZSTD_match4Found_cmov : ZSTD_match4Found_branch; 236 237 DEBUGLOG(5, "ZSTD_compressBlock_fast_generic"); 238 ip0 += (ip0 == prefixStart); 239 { U32 const curr = (U32)(ip0 - base); 240 U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog); 241 U32 const maxRep = curr - windowLow; 242 if (rep_offset2 > maxRep) offsetSaved2 = rep_offset2, rep_offset2 = 0; 243 if (rep_offset1 > maxRep) offsetSaved1 = rep_offset1, rep_offset1 = 0; 244 } 245 246 /* start each op */ 247 _start: /* Requires: ip0 */ 248 249 step = stepSize; 250 nextStep = ip0 + kStepIncr; 251 252 /* calculate positions, ip0 - anchor == 0, so we skip step calc */ 253 ip1 = ip0 + 1; 254 ip2 = ip0 + step; 255 ip3 = ip2 + 1; 256 257 if (ip3 >= ilimit) { 258 goto _cleanup; 259 } 260 261 hash0 = ZSTD_hashPtr(ip0, hlog, mls); 262 hash1 = ZSTD_hashPtr(ip1, hlog, mls); 263 264 matchIdx = hashTable[hash0]; 265 266 do { 267 /* load repcode match for ip[2]*/ 268 const U32 rval = MEM_read32(ip2 - rep_offset1); 269 270 /* write back hash table entry */ 271 current0 = (U32)(ip0 - base); 272 hashTable[hash0] = current0; 273 274 /* check repcode at ip[2] */ 275 if ((MEM_read32(ip2) == rval) & (rep_offset1 > 0)) { 276 ip0 = ip2; 277 match0 = ip0 - rep_offset1; 278 mLength = ip0[-1] == match0[-1]; 279 ip0 -= mLength; 280 match0 -= mLength; 281 offcode = REPCODE1_TO_OFFBASE; 282 mLength += 4; 283 284 /* Write next hash table entry: it's already calculated. 285 * This write is known to be safe because ip1 is before the 286 * repcode (ip2). */ 287 hashTable[hash1] = (U32)(ip1 - base); 288 289 goto _match; 290 } 291 292 if (matchFound(ip0, base + matchIdx, matchIdx, prefixStartIndex)) { 293 /* Write next hash table entry (it's already calculated). 294 * This write is known to be safe because the ip1 == ip0 + 1, 295 * so searching will resume after ip1 */ 296 hashTable[hash1] = (U32)(ip1 - base); 297 298 goto _offset; 299 } 300 301 /* lookup ip[1] */ 302 matchIdx = hashTable[hash1]; 303 304 /* hash ip[2] */ 305 hash0 = hash1; 306 hash1 = ZSTD_hashPtr(ip2, hlog, mls); 307 308 /* advance to next positions */ 309 ip0 = ip1; 310 ip1 = ip2; 311 ip2 = ip3; 312 313 /* write back hash table entry */ 314 current0 = (U32)(ip0 - base); 315 hashTable[hash0] = current0; 316 317 if (matchFound(ip0, base + matchIdx, matchIdx, prefixStartIndex)) { 318 /* Write next hash table entry, since it's already calculated */ 319 if (step <= 4) { 320 /* Avoid writing an index if it's >= position where search will resume. 321 * The minimum possible match has length 4, so search can resume at ip0 + 4. 322 */ 323 hashTable[hash1] = (U32)(ip1 - base); 324 } 325 goto _offset; 326 } 327 328 /* lookup ip[1] */ 329 matchIdx = hashTable[hash1]; 330 331 /* hash ip[2] */ 332 hash0 = hash1; 333 hash1 = ZSTD_hashPtr(ip2, hlog, mls); 334 335 /* advance to next positions */ 336 ip0 = ip1; 337 ip1 = ip2; 338 ip2 = ip0 + step; 339 ip3 = ip1 + step; 340 341 /* calculate step */ 342 if (ip2 >= nextStep) { 343 step++; 344 PREFETCH_L1(ip1 + 64); 345 PREFETCH_L1(ip1 + 128); 346 nextStep += kStepIncr; 347 } 348 } while (ip3 < ilimit); 349 350 _cleanup: 351 /* Note that there are probably still a couple positions one could search. 352 * However, it seems to be a meaningful performance hit to try to search 353 * them. So let's not. */ 354 355 /* When the repcodes are outside of the prefix, we set them to zero before the loop. 356 * When the offsets are still zero, we need to restore them after the block to have a correct 357 * repcode history. If only one offset was invalid, it is easy. The tricky case is when both 358 * offsets were invalid. We need to figure out which offset to refill with. 359 * - If both offsets are zero they are in the same order. 360 * - If both offsets are non-zero, we won't restore the offsets from `offsetSaved[12]`. 361 * - If only one is zero, we need to decide which offset to restore. 362 * - If rep_offset1 is non-zero, then rep_offset2 must be offsetSaved1. 363 * - It is impossible for rep_offset2 to be non-zero. 364 * 365 * So if rep_offset1 started invalid (offsetSaved1 != 0) and became valid (rep_offset1 != 0), then 366 * set rep[0] = rep_offset1 and rep[1] = offsetSaved1. 367 */ 368 offsetSaved2 = ((offsetSaved1 != 0) && (rep_offset1 != 0)) ? offsetSaved1 : offsetSaved2; 369 370 /* save reps for next block */ 371 rep[0] = rep_offset1 ? rep_offset1 : offsetSaved1; 372 rep[1] = rep_offset2 ? rep_offset2 : offsetSaved2; 373 374 /* Return the last literals size */ 375 return (size_t)(iend - anchor); 376 377 _offset: /* Requires: ip0, idx */ 378 379 /* Compute the offset code. */ 380 match0 = base + matchIdx; 381 rep_offset2 = rep_offset1; 382 rep_offset1 = (U32)(ip0-match0); 383 offcode = OFFSET_TO_OFFBASE(rep_offset1); 384 mLength = 4; 385 386 /* Count the backwards match length. */ 387 while (((ip0>anchor) & (match0>prefixStart)) && (ip0[-1] == match0[-1])) { 388 ip0--; 389 match0--; 390 mLength++; 391 } 392 393 _match: /* Requires: ip0, match0, offcode */ 394 395 /* Count the forward length. */ 396 mLength += ZSTD_count(ip0 + mLength, match0 + mLength, iend); 397 398 ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength); 399 400 ip0 += mLength; 401 anchor = ip0; 402 403 /* Fill table and check for immediate repcode. */ 404 if (ip0 <= ilimit) { 405 /* Fill Table */ 406 assert(base+current0+2 > istart); /* check base overflow */ 407 hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2; /* here because current+2 could be > iend-8 */ 408 hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base); 409 410 if (rep_offset2 > 0) { /* rep_offset2==0 means rep_offset2 is invalidated */ 411 while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - rep_offset2)) ) { 412 /* store sequence */ 413 size_t const rLength = ZSTD_count(ip0+4, ip0+4-rep_offset2, iend) + 4; 414 { U32 const tmpOff = rep_offset2; rep_offset2 = rep_offset1; rep_offset1 = tmpOff; } /* swap rep_offset2 <=> rep_offset1 */ 415 hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base); 416 ip0 += rLength; 417 ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, REPCODE1_TO_OFFBASE, rLength); 418 anchor = ip0; 419 continue; /* faster when present (confirmed on gcc-8) ... (?) */ 420 } } } 421 422 goto _start; 423 } 424 425 #define ZSTD_GEN_FAST_FN(dictMode, mml, cmov) \ 426 static size_t ZSTD_compressBlock_fast_##dictMode##_##mml##_##cmov( \ 427 ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], \ 428 void const* src, size_t srcSize) \ 429 { \ 430 return ZSTD_compressBlock_fast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mml, cmov); \ 431 } 432 433 ZSTD_GEN_FAST_FN(noDict, 4, 1) 434 ZSTD_GEN_FAST_FN(noDict, 5, 1) 435 ZSTD_GEN_FAST_FN(noDict, 6, 1) 436 ZSTD_GEN_FAST_FN(noDict, 7, 1) 437 438 ZSTD_GEN_FAST_FN(noDict, 4, 0) 439 ZSTD_GEN_FAST_FN(noDict, 5, 0) 440 ZSTD_GEN_FAST_FN(noDict, 6, 0) 441 ZSTD_GEN_FAST_FN(noDict, 7, 0) 442 443 size_t ZSTD_compressBlock_fast( 444 ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 445 void const* src, size_t srcSize) 446 { 447 U32 const mml = ms->cParams.minMatch; 448 /* use cmov when "candidate in range" branch is likely unpredictable */ 449 int const useCmov = ms->cParams.windowLog < 19; 450 assert(ms->dictMatchState == NULL); 451 if (useCmov) { 452 switch(mml) 453 { 454 default: /* includes case 3 */ 455 case 4 : 456 return ZSTD_compressBlock_fast_noDict_4_1(ms, seqStore, rep, src, srcSize); 457 case 5 : 458 return ZSTD_compressBlock_fast_noDict_5_1(ms, seqStore, rep, src, srcSize); 459 case 6 : 460 return ZSTD_compressBlock_fast_noDict_6_1(ms, seqStore, rep, src, srcSize); 461 case 7 : 462 return ZSTD_compressBlock_fast_noDict_7_1(ms, seqStore, rep, src, srcSize); 463 } 464 } else { 465 /* use a branch instead */ 466 switch(mml) 467 { 468 default: /* includes case 3 */ 469 case 4 : 470 return ZSTD_compressBlock_fast_noDict_4_0(ms, seqStore, rep, src, srcSize); 471 case 5 : 472 return ZSTD_compressBlock_fast_noDict_5_0(ms, seqStore, rep, src, srcSize); 473 case 6 : 474 return ZSTD_compressBlock_fast_noDict_6_0(ms, seqStore, rep, src, srcSize); 475 case 7 : 476 return ZSTD_compressBlock_fast_noDict_7_0(ms, seqStore, rep, src, srcSize); 477 } 478 } 479 } 480 481 FORCE_INLINE_TEMPLATE 482 ZSTD_ALLOW_POINTER_OVERFLOW_ATTR 483 size_t ZSTD_compressBlock_fast_dictMatchState_generic( 484 ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 485 void const* src, size_t srcSize, U32 const mls, U32 const hasStep) 486 { 487 const ZSTD_compressionParameters* const cParams = &ms->cParams; 488 U32* const hashTable = ms->hashTable; 489 U32 const hlog = cParams->hashLog; 490 /* support stepSize of 0 */ 491 U32 const stepSize = cParams->targetLength + !(cParams->targetLength); 492 const BYTE* const base = ms->window.base; 493 const BYTE* const istart = (const BYTE*)src; 494 const BYTE* ip0 = istart; 495 const BYTE* ip1 = ip0 + stepSize; /* we assert below that stepSize >= 1 */ 496 const BYTE* anchor = istart; 497 const U32 prefixStartIndex = ms->window.dictLimit; 498 const BYTE* const prefixStart = base + prefixStartIndex; 499 const BYTE* const iend = istart + srcSize; 500 const BYTE* const ilimit = iend - HASH_READ_SIZE; 501 U32 offset_1=rep[0], offset_2=rep[1]; 502 503 const ZSTD_MatchState_t* const dms = ms->dictMatchState; 504 const ZSTD_compressionParameters* const dictCParams = &dms->cParams ; 505 const U32* const dictHashTable = dms->hashTable; 506 const U32 dictStartIndex = dms->window.dictLimit; 507 const BYTE* const dictBase = dms->window.base; 508 const BYTE* const dictStart = dictBase + dictStartIndex; 509 const BYTE* const dictEnd = dms->window.nextSrc; 510 const U32 dictIndexDelta = prefixStartIndex - (U32)(dictEnd - dictBase); 511 const U32 dictAndPrefixLength = (U32)(istart - prefixStart + dictEnd - dictStart); 512 const U32 dictHBits = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS; 513 514 /* if a dictionary is still attached, it necessarily means that 515 * it is within window size. So we just check it. */ 516 const U32 maxDistance = 1U << cParams->windowLog; 517 const U32 endIndex = (U32)((size_t)(istart - base) + srcSize); 518 assert(endIndex - prefixStartIndex <= maxDistance); 519 (void)maxDistance; (void)endIndex; /* these variables are not used when assert() is disabled */ 520 521 (void)hasStep; /* not currently specialized on whether it's accelerated */ 522 523 /* ensure there will be no underflow 524 * when translating a dict index into a local index */ 525 assert(prefixStartIndex >= (U32)(dictEnd - dictBase)); 526 527 if (ms->prefetchCDictTables) { 528 size_t const hashTableBytes = (((size_t)1) << dictCParams->hashLog) * sizeof(U32); 529 PREFETCH_AREA(dictHashTable, hashTableBytes); 530 } 531 532 /* init */ 533 DEBUGLOG(5, "ZSTD_compressBlock_fast_dictMatchState_generic"); 534 ip0 += (dictAndPrefixLength == 0); 535 /* dictMatchState repCode checks don't currently handle repCode == 0 536 * disabling. */ 537 assert(offset_1 <= dictAndPrefixLength); 538 assert(offset_2 <= dictAndPrefixLength); 539 540 /* Outer search loop */ 541 assert(stepSize >= 1); 542 while (ip1 <= ilimit) { /* repcode check at (ip0 + 1) is safe because ip0 < ip1 */ 543 size_t mLength; 544 size_t hash0 = ZSTD_hashPtr(ip0, hlog, mls); 545 546 size_t const dictHashAndTag0 = ZSTD_hashPtr(ip0, dictHBits, mls); 547 U32 dictMatchIndexAndTag = dictHashTable[dictHashAndTag0 >> ZSTD_SHORT_CACHE_TAG_BITS]; 548 int dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag0); 549 550 U32 matchIndex = hashTable[hash0]; 551 U32 curr = (U32)(ip0 - base); 552 size_t step = stepSize; 553 const size_t kStepIncr = 1 << kSearchStrength; 554 const BYTE* nextStep = ip0 + kStepIncr; 555 556 /* Inner search loop */ 557 while (1) { 558 const BYTE* match = base + matchIndex; 559 const U32 repIndex = curr + 1 - offset_1; 560 const BYTE* repMatch = (repIndex < prefixStartIndex) ? 561 dictBase + (repIndex - dictIndexDelta) : 562 base + repIndex; 563 const size_t hash1 = ZSTD_hashPtr(ip1, hlog, mls); 564 size_t const dictHashAndTag1 = ZSTD_hashPtr(ip1, dictHBits, mls); 565 hashTable[hash0] = curr; /* update hash table */ 566 567 if ((ZSTD_index_overlap_check(prefixStartIndex, repIndex)) 568 && (MEM_read32(repMatch) == MEM_read32(ip0 + 1))) { 569 const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend; 570 mLength = ZSTD_count_2segments(ip0 + 1 + 4, repMatch + 4, iend, repMatchEnd, prefixStart) + 4; 571 ip0++; 572 ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength); 573 break; 574 } 575 576 if (dictTagsMatch) { 577 /* Found a possible dict match */ 578 const U32 dictMatchIndex = dictMatchIndexAndTag >> ZSTD_SHORT_CACHE_TAG_BITS; 579 const BYTE* dictMatch = dictBase + dictMatchIndex; 580 if (dictMatchIndex > dictStartIndex && 581 MEM_read32(dictMatch) == MEM_read32(ip0)) { 582 /* To replicate extDict parse behavior, we only use dict matches when the normal matchIndex is invalid */ 583 if (matchIndex <= prefixStartIndex) { 584 U32 const offset = (U32) (curr - dictMatchIndex - dictIndexDelta); 585 mLength = ZSTD_count_2segments(ip0 + 4, dictMatch + 4, iend, dictEnd, prefixStart) + 4; 586 while (((ip0 > anchor) & (dictMatch > dictStart)) 587 && (ip0[-1] == dictMatch[-1])) { 588 ip0--; 589 dictMatch--; 590 mLength++; 591 } /* catch up */ 592 offset_2 = offset_1; 593 offset_1 = offset; 594 ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength); 595 break; 596 } 597 } 598 } 599 600 if (ZSTD_match4Found_cmov(ip0, match, matchIndex, prefixStartIndex)) { 601 /* found a regular match of size >= 4 */ 602 U32 const offset = (U32) (ip0 - match); 603 mLength = ZSTD_count(ip0 + 4, match + 4, iend) + 4; 604 while (((ip0 > anchor) & (match > prefixStart)) 605 && (ip0[-1] == match[-1])) { 606 ip0--; 607 match--; 608 mLength++; 609 } /* catch up */ 610 offset_2 = offset_1; 611 offset_1 = offset; 612 ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength); 613 break; 614 } 615 616 /* Prepare for next iteration */ 617 dictMatchIndexAndTag = dictHashTable[dictHashAndTag1 >> ZSTD_SHORT_CACHE_TAG_BITS]; 618 dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag1); 619 matchIndex = hashTable[hash1]; 620 621 if (ip1 >= nextStep) { 622 step++; 623 nextStep += kStepIncr; 624 } 625 ip0 = ip1; 626 ip1 = ip1 + step; 627 if (ip1 > ilimit) goto _cleanup; 628 629 curr = (U32)(ip0 - base); 630 hash0 = hash1; 631 } /* end inner search loop */ 632 633 /* match found */ 634 assert(mLength); 635 ip0 += mLength; 636 anchor = ip0; 637 638 if (ip0 <= ilimit) { 639 /* Fill Table */ 640 assert(base+curr+2 > istart); /* check base overflow */ 641 hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2; /* here because curr+2 could be > iend-8 */ 642 hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base); 643 644 /* check immediate repcode */ 645 while (ip0 <= ilimit) { 646 U32 const current2 = (U32)(ip0-base); 647 U32 const repIndex2 = current2 - offset_2; 648 const BYTE* repMatch2 = repIndex2 < prefixStartIndex ? 649 dictBase - dictIndexDelta + repIndex2 : 650 base + repIndex2; 651 if ( (ZSTD_index_overlap_check(prefixStartIndex, repIndex2)) 652 && (MEM_read32(repMatch2) == MEM_read32(ip0))) { 653 const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend; 654 size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4; 655 U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; /* swap offset_2 <=> offset_1 */ 656 ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2); 657 hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = current2; 658 ip0 += repLength2; 659 anchor = ip0; 660 continue; 661 } 662 break; 663 } 664 } 665 666 /* Prepare for next iteration */ 667 assert(ip0 == anchor); 668 ip1 = ip0 + stepSize; 669 } 670 671 _cleanup: 672 /* save reps for next block */ 673 rep[0] = offset_1; 674 rep[1] = offset_2; 675 676 /* Return the last literals size */ 677 return (size_t)(iend - anchor); 678 } 679 680 681 ZSTD_GEN_FAST_FN(dictMatchState, 4, 0) 682 ZSTD_GEN_FAST_FN(dictMatchState, 5, 0) 683 ZSTD_GEN_FAST_FN(dictMatchState, 6, 0) 684 ZSTD_GEN_FAST_FN(dictMatchState, 7, 0) 685 686 size_t ZSTD_compressBlock_fast_dictMatchState( 687 ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 688 void const* src, size_t srcSize) 689 { 690 U32 const mls = ms->cParams.minMatch; 691 assert(ms->dictMatchState != NULL); 692 switch(mls) 693 { 694 default: /* includes case 3 */ 695 case 4 : 696 return ZSTD_compressBlock_fast_dictMatchState_4_0(ms, seqStore, rep, src, srcSize); 697 case 5 : 698 return ZSTD_compressBlock_fast_dictMatchState_5_0(ms, seqStore, rep, src, srcSize); 699 case 6 : 700 return ZSTD_compressBlock_fast_dictMatchState_6_0(ms, seqStore, rep, src, srcSize); 701 case 7 : 702 return ZSTD_compressBlock_fast_dictMatchState_7_0(ms, seqStore, rep, src, srcSize); 703 } 704 } 705 706 707 static 708 ZSTD_ALLOW_POINTER_OVERFLOW_ATTR 709 size_t ZSTD_compressBlock_fast_extDict_generic( 710 ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 711 void const* src, size_t srcSize, U32 const mls, U32 const hasStep) 712 { 713 const ZSTD_compressionParameters* const cParams = &ms->cParams; 714 U32* const hashTable = ms->hashTable; 715 U32 const hlog = cParams->hashLog; 716 /* support stepSize of 0 */ 717 size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1; 718 const BYTE* const base = ms->window.base; 719 const BYTE* const dictBase = ms->window.dictBase; 720 const BYTE* const istart = (const BYTE*)src; 721 const BYTE* anchor = istart; 722 const U32 endIndex = (U32)((size_t)(istart - base) + srcSize); 723 const U32 lowLimit = ZSTD_getLowestMatchIndex(ms, endIndex, cParams->windowLog); 724 const U32 dictStartIndex = lowLimit; 725 const BYTE* const dictStart = dictBase + dictStartIndex; 726 const U32 dictLimit = ms->window.dictLimit; 727 const U32 prefixStartIndex = dictLimit < lowLimit ? lowLimit : dictLimit; 728 const BYTE* const prefixStart = base + prefixStartIndex; 729 const BYTE* const dictEnd = dictBase + prefixStartIndex; 730 const BYTE* const iend = istart + srcSize; 731 const BYTE* const ilimit = iend - 8; 732 U32 offset_1=rep[0], offset_2=rep[1]; 733 U32 offsetSaved1 = 0, offsetSaved2 = 0; 734 735 const BYTE* ip0 = istart; 736 const BYTE* ip1; 737 const BYTE* ip2; 738 const BYTE* ip3; 739 U32 current0; 740 741 742 size_t hash0; /* hash for ip0 */ 743 size_t hash1; /* hash for ip1 */ 744 U32 idx; /* match idx for ip0 */ 745 const BYTE* idxBase; /* base pointer for idx */ 746 747 U32 offcode; 748 const BYTE* match0; 749 size_t mLength; 750 const BYTE* matchEnd = 0; /* initialize to avoid warning, assert != 0 later */ 751 752 size_t step; 753 const BYTE* nextStep; 754 const size_t kStepIncr = (1 << (kSearchStrength - 1)); 755 756 (void)hasStep; /* not currently specialized on whether it's accelerated */ 757 758 DEBUGLOG(5, "ZSTD_compressBlock_fast_extDict_generic (offset_1=%u)", offset_1); 759 760 /* switch to "regular" variant if extDict is invalidated due to maxDistance */ 761 if (prefixStartIndex == dictStartIndex) 762 return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize); 763 764 { U32 const curr = (U32)(ip0 - base); 765 U32 const maxRep = curr - dictStartIndex; 766 if (offset_2 >= maxRep) offsetSaved2 = offset_2, offset_2 = 0; 767 if (offset_1 >= maxRep) offsetSaved1 = offset_1, offset_1 = 0; 768 } 769 770 /* start each op */ 771 _start: /* Requires: ip0 */ 772 773 step = stepSize; 774 nextStep = ip0 + kStepIncr; 775 776 /* calculate positions, ip0 - anchor == 0, so we skip step calc */ 777 ip1 = ip0 + 1; 778 ip2 = ip0 + step; 779 ip3 = ip2 + 1; 780 781 if (ip3 >= ilimit) { 782 goto _cleanup; 783 } 784 785 hash0 = ZSTD_hashPtr(ip0, hlog, mls); 786 hash1 = ZSTD_hashPtr(ip1, hlog, mls); 787 788 idx = hashTable[hash0]; 789 idxBase = idx < prefixStartIndex ? dictBase : base; 790 791 do { 792 { /* load repcode match for ip[2] */ 793 U32 const current2 = (U32)(ip2 - base); 794 U32 const repIndex = current2 - offset_1; 795 const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base; 796 U32 rval; 797 if ( ((U32)(prefixStartIndex - repIndex) >= 4) /* intentional underflow */ 798 & (offset_1 > 0) ) { 799 rval = MEM_read32(repBase + repIndex); 800 } else { 801 rval = MEM_read32(ip2) ^ 1; /* guaranteed to not match. */ 802 } 803 804 /* write back hash table entry */ 805 current0 = (U32)(ip0 - base); 806 hashTable[hash0] = current0; 807 808 /* check repcode at ip[2] */ 809 if (MEM_read32(ip2) == rval) { 810 ip0 = ip2; 811 match0 = repBase + repIndex; 812 matchEnd = repIndex < prefixStartIndex ? dictEnd : iend; 813 assert((match0 != prefixStart) & (match0 != dictStart)); 814 mLength = ip0[-1] == match0[-1]; 815 ip0 -= mLength; 816 match0 -= mLength; 817 offcode = REPCODE1_TO_OFFBASE; 818 mLength += 4; 819 goto _match; 820 } } 821 822 { /* load match for ip[0] */ 823 U32 const mval = idx >= dictStartIndex ? 824 MEM_read32(idxBase + idx) : 825 MEM_read32(ip0) ^ 1; /* guaranteed not to match */ 826 827 /* check match at ip[0] */ 828 if (MEM_read32(ip0) == mval) { 829 /* found a match! */ 830 goto _offset; 831 } } 832 833 /* lookup ip[1] */ 834 idx = hashTable[hash1]; 835 idxBase = idx < prefixStartIndex ? dictBase : base; 836 837 /* hash ip[2] */ 838 hash0 = hash1; 839 hash1 = ZSTD_hashPtr(ip2, hlog, mls); 840 841 /* advance to next positions */ 842 ip0 = ip1; 843 ip1 = ip2; 844 ip2 = ip3; 845 846 /* write back hash table entry */ 847 current0 = (U32)(ip0 - base); 848 hashTable[hash0] = current0; 849 850 { /* load match for ip[0] */ 851 U32 const mval = idx >= dictStartIndex ? 852 MEM_read32(idxBase + idx) : 853 MEM_read32(ip0) ^ 1; /* guaranteed not to match */ 854 855 /* check match at ip[0] */ 856 if (MEM_read32(ip0) == mval) { 857 /* found a match! */ 858 goto _offset; 859 } } 860 861 /* lookup ip[1] */ 862 idx = hashTable[hash1]; 863 idxBase = idx < prefixStartIndex ? dictBase : base; 864 865 /* hash ip[2] */ 866 hash0 = hash1; 867 hash1 = ZSTD_hashPtr(ip2, hlog, mls); 868 869 /* advance to next positions */ 870 ip0 = ip1; 871 ip1 = ip2; 872 ip2 = ip0 + step; 873 ip3 = ip1 + step; 874 875 /* calculate step */ 876 if (ip2 >= nextStep) { 877 step++; 878 PREFETCH_L1(ip1 + 64); 879 PREFETCH_L1(ip1 + 128); 880 nextStep += kStepIncr; 881 } 882 } while (ip3 < ilimit); 883 884 _cleanup: 885 /* Note that there are probably still a couple positions we could search. 886 * However, it seems to be a meaningful performance hit to try to search 887 * them. So let's not. */ 888 889 /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0), 890 * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */ 891 offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2; 892 893 /* save reps for next block */ 894 rep[0] = offset_1 ? offset_1 : offsetSaved1; 895 rep[1] = offset_2 ? offset_2 : offsetSaved2; 896 897 /* Return the last literals size */ 898 return (size_t)(iend - anchor); 899 900 _offset: /* Requires: ip0, idx, idxBase */ 901 902 /* Compute the offset code. */ 903 { U32 const offset = current0 - idx; 904 const BYTE* const lowMatchPtr = idx < prefixStartIndex ? dictStart : prefixStart; 905 matchEnd = idx < prefixStartIndex ? dictEnd : iend; 906 match0 = idxBase + idx; 907 offset_2 = offset_1; 908 offset_1 = offset; 909 offcode = OFFSET_TO_OFFBASE(offset); 910 mLength = 4; 911 912 /* Count the backwards match length. */ 913 while (((ip0>anchor) & (match0>lowMatchPtr)) && (ip0[-1] == match0[-1])) { 914 ip0--; 915 match0--; 916 mLength++; 917 } } 918 919 _match: /* Requires: ip0, match0, offcode, matchEnd */ 920 921 /* Count the forward length. */ 922 assert(matchEnd != 0); 923 mLength += ZSTD_count_2segments(ip0 + mLength, match0 + mLength, iend, matchEnd, prefixStart); 924 925 ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength); 926 927 ip0 += mLength; 928 anchor = ip0; 929 930 /* write next hash table entry */ 931 if (ip1 < ip0) { 932 hashTable[hash1] = (U32)(ip1 - base); 933 } 934 935 /* Fill table and check for immediate repcode. */ 936 if (ip0 <= ilimit) { 937 /* Fill Table */ 938 assert(base+current0+2 > istart); /* check base overflow */ 939 hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2; /* here because current+2 could be > iend-8 */ 940 hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base); 941 942 while (ip0 <= ilimit) { 943 U32 const repIndex2 = (U32)(ip0-base) - offset_2; 944 const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2; 945 if ( ((ZSTD_index_overlap_check(prefixStartIndex, repIndex2)) & (offset_2 > 0)) 946 && (MEM_read32(repMatch2) == MEM_read32(ip0)) ) { 947 const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend; 948 size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4; 949 { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; } /* swap offset_2 <=> offset_1 */ 950 ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, REPCODE1_TO_OFFBASE, repLength2); 951 hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base); 952 ip0 += repLength2; 953 anchor = ip0; 954 continue; 955 } 956 break; 957 } } 958 959 goto _start; 960 } 961 962 ZSTD_GEN_FAST_FN(extDict, 4, 0) 963 ZSTD_GEN_FAST_FN(extDict, 5, 0) 964 ZSTD_GEN_FAST_FN(extDict, 6, 0) 965 ZSTD_GEN_FAST_FN(extDict, 7, 0) 966 967 size_t ZSTD_compressBlock_fast_extDict( 968 ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM], 969 void const* src, size_t srcSize) 970 { 971 U32 const mls = ms->cParams.minMatch; 972 assert(ms->dictMatchState == NULL); 973 switch(mls) 974 { 975 default: /* includes case 3 */ 976 case 4 : 977 return ZSTD_compressBlock_fast_extDict_4_0(ms, seqStore, rep, src, srcSize); 978 case 5 : 979 return ZSTD_compressBlock_fast_extDict_5_0(ms, seqStore, rep, src, srcSize); 980 case 6 : 981 return ZSTD_compressBlock_fast_extDict_6_0(ms, seqStore, rep, src, srcSize); 982 case 7 : 983 return ZSTD_compressBlock_fast_extDict_7_0(ms, seqStore, rep, src, srcSize); 984 } 985 } 986